Most results can be reproduced when data and code are available—but most papers still do not share them
“Investigating the reproducibility of the social and behavioral sciences,” a paper published today in Nature reports testing the reproducibility of a large sample of findings from the social and behavioral sciences and found room for improvement.
Reproducibility refers to whether the same results can be obtained by re-running the same analyses on the same data. This is distinct from replicability, which refers to testing the same question with new data, and robustness, which refers to testing the same question with alternative analyses using the same data.
Scientists and readers alike routinely scrutinize research conclusions: Are the methods sound? Are the measures valid? Would the results hold up under different assumptions? Yet one foundational expectation is often taken for granted: that the numbers reported in a paper are accurate reflections of the analyses actually conducted.
This new investigation from the DARPA-funded Systematizing Confidence in Open Research and Evidence (SCORE) program shows that this expectation—reproducibility—cannot always be assumed.
The study provides the most comprehensive assessment to date of reproducibility—defined as whether the same results can be obtained by re-running the same analyses on the same data— in the social and behavioral sciences.
This collaborative team of 128 researchers from more than 100 institutions around the world examined 600 quantitative research papers published between 2009 and 2018 in 62 prominent journals spanning business, economics, education, political science, psychology, sociology, and related fields. They assessed:
1. Data and code availability: whether authors made their datasets and analytic code accessible
2. Reproducibility: whether independent analysts could reproduce the reported statistical results when data were available.
Key Findings
Most papers do not share their data or code. Only 24% of papers made their data available, and just 20% shared both data and code. Without access to the data, independent assessment of reproducibility is not possible.
When reproduction was possible, it usually succeeded, but a substantial proportion did not. Among papers where reproduction could be attempted, about 74% of papers were reproduced at least approximately, and 54% were reproduced precisely, meaning the numbers matched exactly.
Sharing data and code makes a big difference. When both data and code were available, reproducibility was very high though still not 100%: 91% of paper were reproduced at least approximately and 77% were reproduced precisely. In contrast, when analysts had to reconstruct datasets from an original source (e.g., retrieving data from the census bureau and reconstructing how the authors reported preparing and then analyzing the data), precise reproducibility dropped to about 11%.
Some fields performed better than others. Political science and economics had substantially higher data availability and reproducibility rates than other disciplines. An exploratory investigation suggests that this may have occurred because journals in these fields are more likely to require data sharing, code sharing, and reproducibility checks.
Transparency is improving across the social and behavioral sciences. The proportion of journals requiring data sharing has increased markedly. From 2018 (the last year of the sample of findings reproduced in this study) to 2025, the percentage of journals in the sample requiring data sharing increased from 27% to 52%. If data sharing is causally contributing to reproducibility success, then replicating this study on a more recent sample may show higher success rates.
What These Results Do—and Do Not—Mean
A lack of shared data does not mean a finding is wrong. It means readers cannot independently verify it, creating uncertainty.
A failure to reproduce does not prove that a finding is incorrect. Reproduction can fail due to incomplete documentation, errors by the analyst conducting the reproduction, differences in computing environments, or ambiguities in description of the data preparation and analysis methods.
Successful reproduction does not guarantee a finding is true. Reproducibility confirms precise reporting—not sound theory, valid measurement, replication, or robustness to alternative analyses. Precise reporting is a basic requirement of credibility: are the reported results the same as what is observed from conducting the reported analysis on the data?
Reproducibility is best understood as a baseline quality check, not a final verdict on scientific truth.
Implications
Scientific progress depends on cumulative evidence. When results cannot be independently verified, confidence in that evidence weakens—slowing learning, complicating policy translation, and eroding trust.
The study shows that there is substantial room for improvement in fundamental research behaviors to demonstrate that reported findings are reported precisely and can be verified independently. Making data and code available allows errors in reporting to be detected and corrected.
Conclusion
When researchers share their data and code, most published findings can be reproduced. When they do not, uncertainty remains. Reproducibility is higher when more of the original research materials are accessible. Follow-up research is needed to clarify the causal relationship between open sharing and reproducibility. Nevertheless, it is plausible that sharing data and code increases reproducibility rates, and many journals in the social and behavioral sciences have adopted associated transparency policies.
As lead author Olivia Miske concluded, “Improving reproducibility is not about mistrusting researchers. It is about recognizing that even careful scientists make mistakes. Openness may be an effective tool for quality control.”
Embargoed until publication date: Currently scheduled to be April 1, 2026, 11a ET, 4p UK